1 | #ifndef NTL_mach_desc__H |
---|
2 | #define NTL_mach_desc__H |
---|
3 | |
---|
4 | |
---|
5 | #define NTL_BITS_PER_LONG (64) |
---|
6 | #define NTL_MAX_LONG (9223372036854775807L) |
---|
7 | #define NTL_MAX_INT (2147483647) |
---|
8 | #define NTL_BITS_PER_INT (32) |
---|
9 | #define NTL_BITS_PER_SIZE_T (64) |
---|
10 | #define NTL_ARITH_RIGHT_SHIFT (1) |
---|
11 | #define NTL_NBITS_MAX (50) |
---|
12 | #define NTL_DOUBLE_PRECISION (53) |
---|
13 | #define NTL_FDOUBLE_PRECISION (((double)(1L<<52))) |
---|
14 | #define NTL_QUAD_FLOAT_SPLIT ((((double)(1L<<27)))+1.0) |
---|
15 | #define NTL_EXT_DOUBLE (0) |
---|
16 | #define NTL_SINGLE_MUL_OK (0) |
---|
17 | #define NTL_DOUBLES_LOW_HIGH (0) |
---|
18 | |
---|
19 | |
---|
20 | |
---|
21 | |
---|
22 | #define NTL_BB_MUL_CODE0 \ |
---|
23 | _ntl_ulong hi, lo, t;\ |
---|
24 | _ntl_ulong A[16];\ |
---|
25 | A[0] = 0;\ |
---|
26 | A[1] = a;\ |
---|
27 | A[2] = A[1] << 1;\ |
---|
28 | A[3] = A[2] ^ A[1];\ |
---|
29 | A[4] = A[2] << 1;\ |
---|
30 | A[5] = A[4] ^ A[1];\ |
---|
31 | A[6] = A[3] << 1;\ |
---|
32 | A[7] = A[6] ^ A[1];\ |
---|
33 | A[8] = A[4] << 1;\ |
---|
34 | A[9] = A[8] ^ A[1];\ |
---|
35 | A[10] = A[5] << 1;\ |
---|
36 | A[11] = A[10] ^ A[1];\ |
---|
37 | A[12] = A[6] << 1;\ |
---|
38 | A[13] = A[12] ^ A[1];\ |
---|
39 | A[14] = A[7] << 1;\ |
---|
40 | A[15] = A[14] ^ A[1];\ |
---|
41 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
42 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
43 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
44 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
45 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
46 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
47 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
48 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
49 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
50 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
51 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
52 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
53 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
54 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
55 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
56 | if (a >> 63) hi ^= ((b & 0xeeeeeeeeeeeeeeeeUL) >> 1);\ |
---|
57 | if ((a >> 62) & 1) hi ^= ((b & 0xccccccccccccccccUL) >> 2);\ |
---|
58 | if ((a >> 61) & 1) hi ^= ((b & 0x8888888888888888UL) >> 3);\ |
---|
59 | c[0] = lo; c[1] = hi;\ |
---|
60 | |
---|
61 | |
---|
62 | |
---|
63 | |
---|
64 | |
---|
65 | #define NTL_BB_MUL_CODE1 \ |
---|
66 | long i;\ |
---|
67 | _ntl_ulong carry = 0, b;\ |
---|
68 | _ntl_ulong hi, lo, t;\ |
---|
69 | _ntl_ulong A[16];\ |
---|
70 | A[0] = 0;\ |
---|
71 | A[1] = a;\ |
---|
72 | A[2] = A[1] << 1;\ |
---|
73 | A[3] = A[2] ^ A[1];\ |
---|
74 | A[4] = A[2] << 1;\ |
---|
75 | A[5] = A[4] ^ A[1];\ |
---|
76 | A[6] = A[3] << 1;\ |
---|
77 | A[7] = A[6] ^ A[1];\ |
---|
78 | A[8] = A[4] << 1;\ |
---|
79 | A[9] = A[8] ^ A[1];\ |
---|
80 | A[10] = A[5] << 1;\ |
---|
81 | A[11] = A[10] ^ A[1];\ |
---|
82 | A[12] = A[6] << 1;\ |
---|
83 | A[13] = A[12] ^ A[1];\ |
---|
84 | A[14] = A[7] << 1;\ |
---|
85 | A[15] = A[14] ^ A[1];\ |
---|
86 | for (i = 0; i < sb; i++) {\ |
---|
87 | b = bp[i];\ |
---|
88 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
89 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
90 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
91 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
92 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
93 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
94 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
95 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
96 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
97 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
98 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
99 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
100 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
101 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
102 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
103 | if (a >> 63) hi ^= ((b & 0xeeeeeeeeeeeeeeeeUL) >> 1);\ |
---|
104 | if ((a >> 62) & 1) hi ^= ((b & 0xccccccccccccccccUL) >> 2);\ |
---|
105 | if ((a >> 61) & 1) hi ^= ((b & 0x8888888888888888UL) >> 3);\ |
---|
106 | cp[i] = carry ^ lo; carry = hi;\ |
---|
107 | }\ |
---|
108 | cp[sb] = carry;\ |
---|
109 | |
---|
110 | |
---|
111 | |
---|
112 | |
---|
113 | |
---|
114 | #define NTL_BB_MUL_CODE2 \ |
---|
115 | long i;\ |
---|
116 | _ntl_ulong carry = 0, b;\ |
---|
117 | _ntl_ulong hi, lo, t;\ |
---|
118 | _ntl_ulong A[16];\ |
---|
119 | A[0] = 0;\ |
---|
120 | A[1] = a;\ |
---|
121 | A[2] = A[1] << 1;\ |
---|
122 | A[3] = A[2] ^ A[1];\ |
---|
123 | A[4] = A[2] << 1;\ |
---|
124 | A[5] = A[4] ^ A[1];\ |
---|
125 | A[6] = A[3] << 1;\ |
---|
126 | A[7] = A[6] ^ A[1];\ |
---|
127 | A[8] = A[4] << 1;\ |
---|
128 | A[9] = A[8] ^ A[1];\ |
---|
129 | A[10] = A[5] << 1;\ |
---|
130 | A[11] = A[10] ^ A[1];\ |
---|
131 | A[12] = A[6] << 1;\ |
---|
132 | A[13] = A[12] ^ A[1];\ |
---|
133 | A[14] = A[7] << 1;\ |
---|
134 | A[15] = A[14] ^ A[1];\ |
---|
135 | for (i = 0; i < sb; i++) {\ |
---|
136 | b = bp[i];\ |
---|
137 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
138 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
139 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
140 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
141 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
142 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
143 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
144 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
145 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
146 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
147 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
148 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
149 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
150 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
151 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
152 | if (a >> 63) hi ^= ((b & 0xeeeeeeeeeeeeeeeeUL) >> 1);\ |
---|
153 | if ((a >> 62) & 1) hi ^= ((b & 0xccccccccccccccccUL) >> 2);\ |
---|
154 | if ((a >> 61) & 1) hi ^= ((b & 0x8888888888888888UL) >> 3);\ |
---|
155 | cp[i] ^= (carry ^ lo); carry = hi;\ |
---|
156 | }\ |
---|
157 | cp[sb] ^= carry;\ |
---|
158 | |
---|
159 | |
---|
160 | |
---|
161 | |
---|
162 | |
---|
163 | #define NTL_SHORT_BB_MUL_CODE1 \ |
---|
164 | long i;\ |
---|
165 | _ntl_ulong carry = 0, b;\ |
---|
166 | _ntl_ulong hi, lo, t;\ |
---|
167 | _ntl_ulong A[16];\ |
---|
168 | A[0] = 0;\ |
---|
169 | A[1] = a;\ |
---|
170 | A[2] = A[1] << 1;\ |
---|
171 | A[3] = A[2] ^ A[1];\ |
---|
172 | A[4] = A[2] << 1;\ |
---|
173 | A[5] = A[4] ^ A[1];\ |
---|
174 | A[6] = A[3] << 1;\ |
---|
175 | A[7] = A[6] ^ A[1];\ |
---|
176 | A[8] = A[4] << 1;\ |
---|
177 | A[9] = A[8] ^ A[1];\ |
---|
178 | A[10] = A[5] << 1;\ |
---|
179 | A[11] = A[10] ^ A[1];\ |
---|
180 | A[12] = A[6] << 1;\ |
---|
181 | A[13] = A[12] ^ A[1];\ |
---|
182 | A[14] = A[7] << 1;\ |
---|
183 | A[15] = A[14] ^ A[1];\ |
---|
184 | for (i = 0; i < sb; i++) {\ |
---|
185 | b = bp[i];\ |
---|
186 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
187 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
188 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
189 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
190 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
191 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
192 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
193 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
194 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
195 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
196 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
197 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
198 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
199 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
200 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
201 | cp[i] = carry ^ lo; carry = hi;\ |
---|
202 | }\ |
---|
203 | cp[sb] = carry;\ |
---|
204 | |
---|
205 | |
---|
206 | |
---|
207 | |
---|
208 | |
---|
209 | #define NTL_HALF_BB_MUL_CODE0 \ |
---|
210 | _ntl_ulong hi, lo, t;\ |
---|
211 | _ntl_ulong A[4];\ |
---|
212 | A[0] = 0;\ |
---|
213 | A[1] = a;\ |
---|
214 | A[2] = A[1] << 1;\ |
---|
215 | A[3] = A[2] ^ A[1];\ |
---|
216 | lo = A[b & 3]; t = A[(b >> 2) & 3]; hi = t >> 62; lo ^= t << 2;\ |
---|
217 | t = A[(b >> 4) & 3]; hi ^= t >> 60; lo ^= t << 4;\ |
---|
218 | t = A[(b >> 6) & 3]; hi ^= t >> 58; lo ^= t << 6;\ |
---|
219 | t = A[(b >> 8) & 3]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
220 | t = A[(b >> 10) & 3]; hi ^= t >> 54; lo ^= t << 10;\ |
---|
221 | t = A[(b >> 12) & 3]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
222 | t = A[(b >> 14) & 3]; hi ^= t >> 50; lo ^= t << 14;\ |
---|
223 | t = A[(b >> 16) & 3]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
224 | t = A[(b >> 18) & 3]; hi ^= t >> 46; lo ^= t << 18;\ |
---|
225 | t = A[(b >> 20) & 3]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
226 | t = A[(b >> 22) & 3]; hi ^= t >> 42; lo ^= t << 22;\ |
---|
227 | t = A[(b >> 24) & 3]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
228 | t = A[(b >> 26) & 3]; hi ^= t >> 38; lo ^= t << 26;\ |
---|
229 | t = A[(b >> 28) & 3]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
230 | t = A[b >> 30]; hi ^= t >> 34; lo ^= t << 30;\ |
---|
231 | if (a >> 63) hi ^= ((b & 0xaaaaaaaaUL) >> 1);\ |
---|
232 | c[0] = lo; c[1] = hi;\ |
---|
233 | |
---|
234 | |
---|
235 | |
---|
236 | |
---|
237 | |
---|
238 | #define NTL_ALT_BB_MUL_CODE0 \ |
---|
239 | _ntl_ulong A[16];\ |
---|
240 | A[0] = 0;\ |
---|
241 | A[1] = a;\ |
---|
242 | A[2] = A[1] << 1;\ |
---|
243 | A[3] = A[2] ^ A[1];\ |
---|
244 | A[4] = A[2] << 1;\ |
---|
245 | A[5] = A[4] ^ A[1];\ |
---|
246 | A[6] = A[3] << 1;\ |
---|
247 | A[7] = A[6] ^ A[1];\ |
---|
248 | A[8] = A[4] << 1;\ |
---|
249 | A[9] = A[8] ^ A[1];\ |
---|
250 | A[10] = A[5] << 1;\ |
---|
251 | A[11] = A[10] ^ A[1];\ |
---|
252 | A[12] = A[6] << 1;\ |
---|
253 | A[13] = A[12] ^ A[1];\ |
---|
254 | A[14] = A[7] << 1;\ |
---|
255 | A[15] = A[14] ^ A[1];\ |
---|
256 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
257 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
258 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
259 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
260 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
261 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
262 | const _ntl_ulong t28 = A[(b >> 28) & 15]; \ |
---|
263 | const _ntl_ulong t32 = A[(b >> 32) & 15]; \ |
---|
264 | const _ntl_ulong t36 = A[(b >> 36) & 15]; \ |
---|
265 | const _ntl_ulong t40 = A[(b >> 40) & 15]; \ |
---|
266 | const _ntl_ulong t44 = A[(b >> 44) & 15]; \ |
---|
267 | const _ntl_ulong t48 = A[(b >> 48) & 15]; \ |
---|
268 | const _ntl_ulong t52 = A[(b >> 52) & 15]; \ |
---|
269 | const _ntl_ulong t56 = A[(b >> 56) & 15]; \ |
---|
270 | const _ntl_ulong t60 = A[b >> 60]; \ |
---|
271 | const _ntl_ulong lo = A[b & 15] \ |
---|
272 | ^ (t4 << 4)\ |
---|
273 | ^ (t8 << 8)\ |
---|
274 | ^ (t12 << 12)\ |
---|
275 | ^ (t16 << 16)\ |
---|
276 | ^ (t20 << 20)\ |
---|
277 | ^ (t24 << 24)\ |
---|
278 | ^ (t28 << 28)\ |
---|
279 | ^ (t32 << 32)\ |
---|
280 | ^ (t36 << 36)\ |
---|
281 | ^ (t40 << 40)\ |
---|
282 | ^ (t44 << 44)\ |
---|
283 | ^ (t48 << 48)\ |
---|
284 | ^ (t52 << 52)\ |
---|
285 | ^ (t56 << 56)\ |
---|
286 | ^ (t60 << 60);\ |
---|
287 | const _ntl_ulong hi = (t4 >> 60)\ |
---|
288 | ^ (t8 >> 56)\ |
---|
289 | ^ (t12 >> 52)\ |
---|
290 | ^ (t16 >> 48)\ |
---|
291 | ^ (t20 >> 44)\ |
---|
292 | ^ (t24 >> 40)\ |
---|
293 | ^ (t28 >> 36)\ |
---|
294 | ^ (t32 >> 32)\ |
---|
295 | ^ (t36 >> 28)\ |
---|
296 | ^ (t40 >> 24)\ |
---|
297 | ^ (t44 >> 20)\ |
---|
298 | ^ (t48 >> 16)\ |
---|
299 | ^ (t52 >> 12)\ |
---|
300 | ^ (t56 >> 8)\ |
---|
301 | ^ (t60 >> 4)\ |
---|
302 | ^ (((b & 0xeeeeeeeeeeeeeeeeUL) >> 1) & (-(a >> 63)))\ |
---|
303 | ^ (((b & 0xccccccccccccccccUL) >> 2) & (-((a >> 62) & 1UL)))\ |
---|
304 | ^ (((b & 0x8888888888888888UL) >> 3) & (-((a >> 61) & 1UL)));\ |
---|
305 | c[0] = lo; c[1] = hi;\ |
---|
306 | |
---|
307 | |
---|
308 | |
---|
309 | |
---|
310 | |
---|
311 | #define NTL_ALT_BB_MUL_CODE1 \ |
---|
312 | long i;\ |
---|
313 | _ntl_ulong carry = 0;\ |
---|
314 | _ntl_ulong A[16];\ |
---|
315 | A[0] = 0;\ |
---|
316 | A[1] = a;\ |
---|
317 | A[2] = A[1] << 1;\ |
---|
318 | A[3] = A[2] ^ A[1];\ |
---|
319 | A[4] = A[2] << 1;\ |
---|
320 | A[5] = A[4] ^ A[1];\ |
---|
321 | A[6] = A[3] << 1;\ |
---|
322 | A[7] = A[6] ^ A[1];\ |
---|
323 | A[8] = A[4] << 1;\ |
---|
324 | A[9] = A[8] ^ A[1];\ |
---|
325 | A[10] = A[5] << 1;\ |
---|
326 | A[11] = A[10] ^ A[1];\ |
---|
327 | A[12] = A[6] << 1;\ |
---|
328 | A[13] = A[12] ^ A[1];\ |
---|
329 | A[14] = A[7] << 1;\ |
---|
330 | A[15] = A[14] ^ A[1];\ |
---|
331 | for (i = 0; i < sb; i++) {\ |
---|
332 | const _ntl_ulong b = bp[i];\ |
---|
333 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
334 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
335 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
336 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
337 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
338 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
339 | const _ntl_ulong t28 = A[(b >> 28) & 15]; \ |
---|
340 | const _ntl_ulong t32 = A[(b >> 32) & 15]; \ |
---|
341 | const _ntl_ulong t36 = A[(b >> 36) & 15]; \ |
---|
342 | const _ntl_ulong t40 = A[(b >> 40) & 15]; \ |
---|
343 | const _ntl_ulong t44 = A[(b >> 44) & 15]; \ |
---|
344 | const _ntl_ulong t48 = A[(b >> 48) & 15]; \ |
---|
345 | const _ntl_ulong t52 = A[(b >> 52) & 15]; \ |
---|
346 | const _ntl_ulong t56 = A[(b >> 56) & 15]; \ |
---|
347 | const _ntl_ulong t60 = A[b >> 60]; \ |
---|
348 | const _ntl_ulong lo = A[b & 15] \ |
---|
349 | ^ (t4 << 4)\ |
---|
350 | ^ (t8 << 8)\ |
---|
351 | ^ (t12 << 12)\ |
---|
352 | ^ (t16 << 16)\ |
---|
353 | ^ (t20 << 20)\ |
---|
354 | ^ (t24 << 24)\ |
---|
355 | ^ (t28 << 28)\ |
---|
356 | ^ (t32 << 32)\ |
---|
357 | ^ (t36 << 36)\ |
---|
358 | ^ (t40 << 40)\ |
---|
359 | ^ (t44 << 44)\ |
---|
360 | ^ (t48 << 48)\ |
---|
361 | ^ (t52 << 52)\ |
---|
362 | ^ (t56 << 56)\ |
---|
363 | ^ (t60 << 60);\ |
---|
364 | const _ntl_ulong hi = (t4 >> 60)\ |
---|
365 | ^ (t8 >> 56)\ |
---|
366 | ^ (t12 >> 52)\ |
---|
367 | ^ (t16 >> 48)\ |
---|
368 | ^ (t20 >> 44)\ |
---|
369 | ^ (t24 >> 40)\ |
---|
370 | ^ (t28 >> 36)\ |
---|
371 | ^ (t32 >> 32)\ |
---|
372 | ^ (t36 >> 28)\ |
---|
373 | ^ (t40 >> 24)\ |
---|
374 | ^ (t44 >> 20)\ |
---|
375 | ^ (t48 >> 16)\ |
---|
376 | ^ (t52 >> 12)\ |
---|
377 | ^ (t56 >> 8)\ |
---|
378 | ^ (t60 >> 4)\ |
---|
379 | ^ (((b & 0xeeeeeeeeeeeeeeeeUL) >> 1) & (-(a >> 63)))\ |
---|
380 | ^ (((b & 0xccccccccccccccccUL) >> 2) & (-((a >> 62) & 1UL)))\ |
---|
381 | ^ (((b & 0x8888888888888888UL) >> 3) & (-((a >> 61) & 1UL)));\ |
---|
382 | cp[i] = carry ^ lo; carry = hi;\ |
---|
383 | }\ |
---|
384 | cp[sb] = carry;\ |
---|
385 | |
---|
386 | |
---|
387 | |
---|
388 | |
---|
389 | |
---|
390 | #define NTL_ALT_BB_MUL_CODE2 \ |
---|
391 | long i;\ |
---|
392 | _ntl_ulong carry = 0;\ |
---|
393 | _ntl_ulong A[16];\ |
---|
394 | A[0] = 0;\ |
---|
395 | A[1] = a;\ |
---|
396 | A[2] = A[1] << 1;\ |
---|
397 | A[3] = A[2] ^ A[1];\ |
---|
398 | A[4] = A[2] << 1;\ |
---|
399 | A[5] = A[4] ^ A[1];\ |
---|
400 | A[6] = A[3] << 1;\ |
---|
401 | A[7] = A[6] ^ A[1];\ |
---|
402 | A[8] = A[4] << 1;\ |
---|
403 | A[9] = A[8] ^ A[1];\ |
---|
404 | A[10] = A[5] << 1;\ |
---|
405 | A[11] = A[10] ^ A[1];\ |
---|
406 | A[12] = A[6] << 1;\ |
---|
407 | A[13] = A[12] ^ A[1];\ |
---|
408 | A[14] = A[7] << 1;\ |
---|
409 | A[15] = A[14] ^ A[1];\ |
---|
410 | for (i = 0; i < sb; i++) {\ |
---|
411 | const _ntl_ulong b = bp[i];\ |
---|
412 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
413 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
414 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
415 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
416 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
417 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
418 | const _ntl_ulong t28 = A[(b >> 28) & 15]; \ |
---|
419 | const _ntl_ulong t32 = A[(b >> 32) & 15]; \ |
---|
420 | const _ntl_ulong t36 = A[(b >> 36) & 15]; \ |
---|
421 | const _ntl_ulong t40 = A[(b >> 40) & 15]; \ |
---|
422 | const _ntl_ulong t44 = A[(b >> 44) & 15]; \ |
---|
423 | const _ntl_ulong t48 = A[(b >> 48) & 15]; \ |
---|
424 | const _ntl_ulong t52 = A[(b >> 52) & 15]; \ |
---|
425 | const _ntl_ulong t56 = A[(b >> 56) & 15]; \ |
---|
426 | const _ntl_ulong t60 = A[b >> 60]; \ |
---|
427 | const _ntl_ulong lo = A[b & 15] \ |
---|
428 | ^ (t4 << 4)\ |
---|
429 | ^ (t8 << 8)\ |
---|
430 | ^ (t12 << 12)\ |
---|
431 | ^ (t16 << 16)\ |
---|
432 | ^ (t20 << 20)\ |
---|
433 | ^ (t24 << 24)\ |
---|
434 | ^ (t28 << 28)\ |
---|
435 | ^ (t32 << 32)\ |
---|
436 | ^ (t36 << 36)\ |
---|
437 | ^ (t40 << 40)\ |
---|
438 | ^ (t44 << 44)\ |
---|
439 | ^ (t48 << 48)\ |
---|
440 | ^ (t52 << 52)\ |
---|
441 | ^ (t56 << 56)\ |
---|
442 | ^ (t60 << 60);\ |
---|
443 | const _ntl_ulong hi = (t4 >> 60)\ |
---|
444 | ^ (t8 >> 56)\ |
---|
445 | ^ (t12 >> 52)\ |
---|
446 | ^ (t16 >> 48)\ |
---|
447 | ^ (t20 >> 44)\ |
---|
448 | ^ (t24 >> 40)\ |
---|
449 | ^ (t28 >> 36)\ |
---|
450 | ^ (t32 >> 32)\ |
---|
451 | ^ (t36 >> 28)\ |
---|
452 | ^ (t40 >> 24)\ |
---|
453 | ^ (t44 >> 20)\ |
---|
454 | ^ (t48 >> 16)\ |
---|
455 | ^ (t52 >> 12)\ |
---|
456 | ^ (t56 >> 8)\ |
---|
457 | ^ (t60 >> 4)\ |
---|
458 | ^ (((b & 0xeeeeeeeeeeeeeeeeUL) >> 1) & (-(a >> 63)))\ |
---|
459 | ^ (((b & 0xccccccccccccccccUL) >> 2) & (-((a >> 62) & 1UL)))\ |
---|
460 | ^ (((b & 0x8888888888888888UL) >> 3) & (-((a >> 61) & 1UL)));\ |
---|
461 | cp[i] ^= (carry ^ lo); carry = hi;\ |
---|
462 | }\ |
---|
463 | cp[sb] ^= carry;\ |
---|
464 | |
---|
465 | |
---|
466 | |
---|
467 | |
---|
468 | |
---|
469 | #define NTL_ALT_SHORT_BB_MUL_CODE1 \ |
---|
470 | long i;\ |
---|
471 | _ntl_ulong carry = 0;\ |
---|
472 | _ntl_ulong A[16];\ |
---|
473 | A[0] = 0;\ |
---|
474 | A[1] = a;\ |
---|
475 | A[2] = A[1] << 1;\ |
---|
476 | A[3] = A[2] ^ A[1];\ |
---|
477 | A[4] = A[2] << 1;\ |
---|
478 | A[5] = A[4] ^ A[1];\ |
---|
479 | A[6] = A[3] << 1;\ |
---|
480 | A[7] = A[6] ^ A[1];\ |
---|
481 | A[8] = A[4] << 1;\ |
---|
482 | A[9] = A[8] ^ A[1];\ |
---|
483 | A[10] = A[5] << 1;\ |
---|
484 | A[11] = A[10] ^ A[1];\ |
---|
485 | A[12] = A[6] << 1;\ |
---|
486 | A[13] = A[12] ^ A[1];\ |
---|
487 | A[14] = A[7] << 1;\ |
---|
488 | A[15] = A[14] ^ A[1];\ |
---|
489 | for (i = 0; i < sb; i++) {\ |
---|
490 | const _ntl_ulong b = bp[i];\ |
---|
491 | const _ntl_ulong t4 = A[(b >> 4) & 15]; \ |
---|
492 | const _ntl_ulong t8 = A[(b >> 8) & 15]; \ |
---|
493 | const _ntl_ulong t12 = A[(b >> 12) & 15]; \ |
---|
494 | const _ntl_ulong t16 = A[(b >> 16) & 15]; \ |
---|
495 | const _ntl_ulong t20 = A[(b >> 20) & 15]; \ |
---|
496 | const _ntl_ulong t24 = A[(b >> 24) & 15]; \ |
---|
497 | const _ntl_ulong t28 = A[(b >> 28) & 15]; \ |
---|
498 | const _ntl_ulong t32 = A[(b >> 32) & 15]; \ |
---|
499 | const _ntl_ulong t36 = A[(b >> 36) & 15]; \ |
---|
500 | const _ntl_ulong t40 = A[(b >> 40) & 15]; \ |
---|
501 | const _ntl_ulong t44 = A[(b >> 44) & 15]; \ |
---|
502 | const _ntl_ulong t48 = A[(b >> 48) & 15]; \ |
---|
503 | const _ntl_ulong t52 = A[(b >> 52) & 15]; \ |
---|
504 | const _ntl_ulong t56 = A[(b >> 56) & 15]; \ |
---|
505 | const _ntl_ulong t60 = A[b >> 60]; \ |
---|
506 | const _ntl_ulong lo = A[b & 15] \ |
---|
507 | ^ (t4 << 4)\ |
---|
508 | ^ (t8 << 8)\ |
---|
509 | ^ (t12 << 12)\ |
---|
510 | ^ (t16 << 16)\ |
---|
511 | ^ (t20 << 20)\ |
---|
512 | ^ (t24 << 24)\ |
---|
513 | ^ (t28 << 28)\ |
---|
514 | ^ (t32 << 32)\ |
---|
515 | ^ (t36 << 36)\ |
---|
516 | ^ (t40 << 40)\ |
---|
517 | ^ (t44 << 44)\ |
---|
518 | ^ (t48 << 48)\ |
---|
519 | ^ (t52 << 52)\ |
---|
520 | ^ (t56 << 56)\ |
---|
521 | ^ (t60 << 60);\ |
---|
522 | const _ntl_ulong hi = (t4 >> 60)\ |
---|
523 | ^ (t8 >> 56)\ |
---|
524 | ^ (t12 >> 52)\ |
---|
525 | ^ (t16 >> 48)\ |
---|
526 | ^ (t20 >> 44)\ |
---|
527 | ^ (t24 >> 40)\ |
---|
528 | ^ (t28 >> 36)\ |
---|
529 | ^ (t32 >> 32)\ |
---|
530 | ^ (t36 >> 28)\ |
---|
531 | ^ (t40 >> 24)\ |
---|
532 | ^ (t44 >> 20)\ |
---|
533 | ^ (t48 >> 16)\ |
---|
534 | ^ (t52 >> 12)\ |
---|
535 | ^ (t56 >> 8)\ |
---|
536 | ^ (t60 >> 4);\ |
---|
537 | cp[i] = carry ^ lo; carry = hi;\ |
---|
538 | }\ |
---|
539 | cp[sb] = carry;\ |
---|
540 | |
---|
541 | |
---|
542 | |
---|
543 | |
---|
544 | |
---|
545 | #define NTL_ALT_HALF_BB_MUL_CODE0 \ |
---|
546 | _ntl_ulong A[4];\ |
---|
547 | A[0] = 0;\ |
---|
548 | A[1] = a;\ |
---|
549 | A[2] = A[1] << 1;\ |
---|
550 | A[3] = A[2] ^ A[1];\ |
---|
551 | const _ntl_ulong t2 = A[(b >> 2) & 3]; \ |
---|
552 | const _ntl_ulong t4 = A[(b >> 4) & 3]; \ |
---|
553 | const _ntl_ulong t6 = A[(b >> 6) & 3]; \ |
---|
554 | const _ntl_ulong t8 = A[(b >> 8) & 3]; \ |
---|
555 | const _ntl_ulong t10 = A[(b >> 10) & 3]; \ |
---|
556 | const _ntl_ulong t12 = A[(b >> 12) & 3]; \ |
---|
557 | const _ntl_ulong t14 = A[(b >> 14) & 3]; \ |
---|
558 | const _ntl_ulong t16 = A[(b >> 16) & 3]; \ |
---|
559 | const _ntl_ulong t18 = A[(b >> 18) & 3]; \ |
---|
560 | const _ntl_ulong t20 = A[(b >> 20) & 3]; \ |
---|
561 | const _ntl_ulong t22 = A[(b >> 22) & 3]; \ |
---|
562 | const _ntl_ulong t24 = A[(b >> 24) & 3]; \ |
---|
563 | const _ntl_ulong t26 = A[(b >> 26) & 3]; \ |
---|
564 | const _ntl_ulong t28 = A[(b >> 28) & 3]; \ |
---|
565 | const _ntl_ulong t30 = A[b >> 30]; \ |
---|
566 | const _ntl_ulong lo = A[b & 3] \ |
---|
567 | ^ (t2 << 2)\ |
---|
568 | ^ (t4 << 4)\ |
---|
569 | ^ (t6 << 6)\ |
---|
570 | ^ (t8 << 8)\ |
---|
571 | ^ (t10 << 10)\ |
---|
572 | ^ (t12 << 12)\ |
---|
573 | ^ (t14 << 14)\ |
---|
574 | ^ (t16 << 16)\ |
---|
575 | ^ (t18 << 18)\ |
---|
576 | ^ (t20 << 20)\ |
---|
577 | ^ (t22 << 22)\ |
---|
578 | ^ (t24 << 24)\ |
---|
579 | ^ (t26 << 26)\ |
---|
580 | ^ (t28 << 28)\ |
---|
581 | ^ (t30 << 30);\ |
---|
582 | const _ntl_ulong hi = (t2 >> 62)\ |
---|
583 | ^ (t4 >> 60)\ |
---|
584 | ^ (t6 >> 58)\ |
---|
585 | ^ (t8 >> 56)\ |
---|
586 | ^ (t10 >> 54)\ |
---|
587 | ^ (t12 >> 52)\ |
---|
588 | ^ (t14 >> 50)\ |
---|
589 | ^ (t16 >> 48)\ |
---|
590 | ^ (t18 >> 46)\ |
---|
591 | ^ (t20 >> 44)\ |
---|
592 | ^ (t22 >> 42)\ |
---|
593 | ^ (t24 >> 40)\ |
---|
594 | ^ (t26 >> 38)\ |
---|
595 | ^ (t28 >> 36)\ |
---|
596 | ^ (t30 >> 34)\ |
---|
597 | ^ (((b & 0xaaaaaaaaUL) >> 1) & (-(a >> 63)));\ |
---|
598 | c[0] = lo; c[1] = hi;\ |
---|
599 | |
---|
600 | |
---|
601 | |
---|
602 | |
---|
603 | |
---|
604 | #define NTL_ALT1_BB_MUL_CODE0 \ |
---|
605 | _ntl_ulong hi, lo, t;\ |
---|
606 | _ntl_ulong A[16];\ |
---|
607 | A[0] = 0;\ |
---|
608 | A[1] = a;\ |
---|
609 | A[2] = A[1] << 1;\ |
---|
610 | A[3] = A[2] ^ A[1];\ |
---|
611 | A[4] = A[2] << 1;\ |
---|
612 | A[5] = A[4] ^ A[1];\ |
---|
613 | A[6] = A[3] << 1;\ |
---|
614 | A[7] = A[6] ^ A[1];\ |
---|
615 | A[8] = A[4] << 1;\ |
---|
616 | A[9] = A[8] ^ A[1];\ |
---|
617 | A[10] = A[5] << 1;\ |
---|
618 | A[11] = A[10] ^ A[1];\ |
---|
619 | A[12] = A[6] << 1;\ |
---|
620 | A[13] = A[12] ^ A[1];\ |
---|
621 | A[14] = A[7] << 1;\ |
---|
622 | A[15] = A[14] ^ A[1];\ |
---|
623 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
624 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
625 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
626 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
627 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
628 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
629 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
630 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
631 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
632 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
633 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
634 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
635 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
636 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
637 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
638 | hi ^= (((b & 0xeeeeeeeeeeeeeeeeUL) >> 1) & (-(a >> 63)))\ |
---|
639 | ^ (((b & 0xccccccccccccccccUL) >> 2) & (-((a >> 62) & 1UL)))\ |
---|
640 | ^ (((b & 0x8888888888888888UL) >> 3) & (-((a >> 61) & 1UL)));\ |
---|
641 | c[0] = lo; c[1] = hi;\ |
---|
642 | |
---|
643 | |
---|
644 | |
---|
645 | |
---|
646 | |
---|
647 | #define NTL_ALT1_BB_MUL_CODE1 \ |
---|
648 | long i;\ |
---|
649 | _ntl_ulong carry = 0, b;\ |
---|
650 | _ntl_ulong hi, lo, t;\ |
---|
651 | _ntl_ulong A[16];\ |
---|
652 | A[0] = 0;\ |
---|
653 | A[1] = a;\ |
---|
654 | A[2] = A[1] << 1;\ |
---|
655 | A[3] = A[2] ^ A[1];\ |
---|
656 | A[4] = A[2] << 1;\ |
---|
657 | A[5] = A[4] ^ A[1];\ |
---|
658 | A[6] = A[3] << 1;\ |
---|
659 | A[7] = A[6] ^ A[1];\ |
---|
660 | A[8] = A[4] << 1;\ |
---|
661 | A[9] = A[8] ^ A[1];\ |
---|
662 | A[10] = A[5] << 1;\ |
---|
663 | A[11] = A[10] ^ A[1];\ |
---|
664 | A[12] = A[6] << 1;\ |
---|
665 | A[13] = A[12] ^ A[1];\ |
---|
666 | A[14] = A[7] << 1;\ |
---|
667 | A[15] = A[14] ^ A[1];\ |
---|
668 | for (i = 0; i < sb; i++) {\ |
---|
669 | b = bp[i];\ |
---|
670 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
671 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
672 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
673 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
674 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
675 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
676 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
677 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
678 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
679 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
680 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
681 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
682 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
683 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
684 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
685 | hi ^= (((b & 0xeeeeeeeeeeeeeeeeUL) >> 1) & (-(a >> 63)))\ |
---|
686 | ^ (((b & 0xccccccccccccccccUL) >> 2) & (-((a >> 62) & 1UL)))\ |
---|
687 | ^ (((b & 0x8888888888888888UL) >> 3) & (-((a >> 61) & 1UL)));\ |
---|
688 | cp[i] = carry ^ lo; carry = hi;\ |
---|
689 | }\ |
---|
690 | cp[sb] = carry;\ |
---|
691 | |
---|
692 | |
---|
693 | |
---|
694 | |
---|
695 | |
---|
696 | #define NTL_ALT1_BB_MUL_CODE2 \ |
---|
697 | long i;\ |
---|
698 | _ntl_ulong carry = 0, b;\ |
---|
699 | _ntl_ulong hi, lo, t;\ |
---|
700 | _ntl_ulong A[16];\ |
---|
701 | A[0] = 0;\ |
---|
702 | A[1] = a;\ |
---|
703 | A[2] = A[1] << 1;\ |
---|
704 | A[3] = A[2] ^ A[1];\ |
---|
705 | A[4] = A[2] << 1;\ |
---|
706 | A[5] = A[4] ^ A[1];\ |
---|
707 | A[6] = A[3] << 1;\ |
---|
708 | A[7] = A[6] ^ A[1];\ |
---|
709 | A[8] = A[4] << 1;\ |
---|
710 | A[9] = A[8] ^ A[1];\ |
---|
711 | A[10] = A[5] << 1;\ |
---|
712 | A[11] = A[10] ^ A[1];\ |
---|
713 | A[12] = A[6] << 1;\ |
---|
714 | A[13] = A[12] ^ A[1];\ |
---|
715 | A[14] = A[7] << 1;\ |
---|
716 | A[15] = A[14] ^ A[1];\ |
---|
717 | for (i = 0; i < sb; i++) {\ |
---|
718 | b = bp[i];\ |
---|
719 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
720 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
721 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
722 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
723 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
724 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
725 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
726 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
727 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
728 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
729 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
730 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
731 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
732 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
733 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
734 | hi ^= (((b & 0xeeeeeeeeeeeeeeeeUL) >> 1) & (-(a >> 63)))\ |
---|
735 | ^ (((b & 0xccccccccccccccccUL) >> 2) & (-((a >> 62) & 1UL)))\ |
---|
736 | ^ (((b & 0x8888888888888888UL) >> 3) & (-((a >> 61) & 1UL)));\ |
---|
737 | cp[i] ^= (carry ^ lo); carry = hi;\ |
---|
738 | }\ |
---|
739 | cp[sb] ^= carry;\ |
---|
740 | |
---|
741 | |
---|
742 | |
---|
743 | |
---|
744 | |
---|
745 | #define NTL_ALT1_SHORT_BB_MUL_CODE1 \ |
---|
746 | long i;\ |
---|
747 | _ntl_ulong carry = 0, b;\ |
---|
748 | _ntl_ulong hi, lo, t;\ |
---|
749 | _ntl_ulong A[16];\ |
---|
750 | A[0] = 0;\ |
---|
751 | A[1] = a;\ |
---|
752 | A[2] = A[1] << 1;\ |
---|
753 | A[3] = A[2] ^ A[1];\ |
---|
754 | A[4] = A[2] << 1;\ |
---|
755 | A[5] = A[4] ^ A[1];\ |
---|
756 | A[6] = A[3] << 1;\ |
---|
757 | A[7] = A[6] ^ A[1];\ |
---|
758 | A[8] = A[4] << 1;\ |
---|
759 | A[9] = A[8] ^ A[1];\ |
---|
760 | A[10] = A[5] << 1;\ |
---|
761 | A[11] = A[10] ^ A[1];\ |
---|
762 | A[12] = A[6] << 1;\ |
---|
763 | A[13] = A[12] ^ A[1];\ |
---|
764 | A[14] = A[7] << 1;\ |
---|
765 | A[15] = A[14] ^ A[1];\ |
---|
766 | for (i = 0; i < sb; i++) {\ |
---|
767 | b = bp[i];\ |
---|
768 | lo = A[b & 15]; t = A[(b >> 4) & 15]; hi = t >> 60; lo ^= t << 4;\ |
---|
769 | t = A[(b >> 8) & 15]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
770 | t = A[(b >> 12) & 15]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
771 | t = A[(b >> 16) & 15]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
772 | t = A[(b >> 20) & 15]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
773 | t = A[(b >> 24) & 15]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
774 | t = A[(b >> 28) & 15]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
775 | t = A[(b >> 32) & 15]; hi ^= t >> 32; lo ^= t << 32;\ |
---|
776 | t = A[(b >> 36) & 15]; hi ^= t >> 28; lo ^= t << 36;\ |
---|
777 | t = A[(b >> 40) & 15]; hi ^= t >> 24; lo ^= t << 40;\ |
---|
778 | t = A[(b >> 44) & 15]; hi ^= t >> 20; lo ^= t << 44;\ |
---|
779 | t = A[(b >> 48) & 15]; hi ^= t >> 16; lo ^= t << 48;\ |
---|
780 | t = A[(b >> 52) & 15]; hi ^= t >> 12; lo ^= t << 52;\ |
---|
781 | t = A[(b >> 56) & 15]; hi ^= t >> 8; lo ^= t << 56;\ |
---|
782 | t = A[b >> 60]; hi ^= t >> 4; lo ^= t << 60;\ |
---|
783 | cp[i] = carry ^ lo; carry = hi;\ |
---|
784 | }\ |
---|
785 | cp[sb] = carry;\ |
---|
786 | |
---|
787 | |
---|
788 | |
---|
789 | |
---|
790 | |
---|
791 | #define NTL_ALT1_HALF_BB_MUL_CODE0 \ |
---|
792 | _ntl_ulong hi, lo, t;\ |
---|
793 | _ntl_ulong A[4];\ |
---|
794 | A[0] = 0;\ |
---|
795 | A[1] = a;\ |
---|
796 | A[2] = A[1] << 1;\ |
---|
797 | A[3] = A[2] ^ A[1];\ |
---|
798 | lo = A[b & 3]; t = A[(b >> 2) & 3]; hi = t >> 62; lo ^= t << 2;\ |
---|
799 | t = A[(b >> 4) & 3]; hi ^= t >> 60; lo ^= t << 4;\ |
---|
800 | t = A[(b >> 6) & 3]; hi ^= t >> 58; lo ^= t << 6;\ |
---|
801 | t = A[(b >> 8) & 3]; hi ^= t >> 56; lo ^= t << 8;\ |
---|
802 | t = A[(b >> 10) & 3]; hi ^= t >> 54; lo ^= t << 10;\ |
---|
803 | t = A[(b >> 12) & 3]; hi ^= t >> 52; lo ^= t << 12;\ |
---|
804 | t = A[(b >> 14) & 3]; hi ^= t >> 50; lo ^= t << 14;\ |
---|
805 | t = A[(b >> 16) & 3]; hi ^= t >> 48; lo ^= t << 16;\ |
---|
806 | t = A[(b >> 18) & 3]; hi ^= t >> 46; lo ^= t << 18;\ |
---|
807 | t = A[(b >> 20) & 3]; hi ^= t >> 44; lo ^= t << 20;\ |
---|
808 | t = A[(b >> 22) & 3]; hi ^= t >> 42; lo ^= t << 22;\ |
---|
809 | t = A[(b >> 24) & 3]; hi ^= t >> 40; lo ^= t << 24;\ |
---|
810 | t = A[(b >> 26) & 3]; hi ^= t >> 38; lo ^= t << 26;\ |
---|
811 | t = A[(b >> 28) & 3]; hi ^= t >> 36; lo ^= t << 28;\ |
---|
812 | t = A[b >> 30]; hi ^= t >> 34; lo ^= t << 30;\ |
---|
813 | hi ^= (((b & 0xaaaaaaaaUL) >> 1) & (-(a >> 63)));\ |
---|
814 | c[0] = lo; c[1] = hi;\ |
---|
815 | |
---|
816 | |
---|
817 | |
---|
818 | #define NTL_BB_MUL1_BITS (4) |
---|
819 | |
---|
820 | |
---|
821 | |
---|
822 | #define NTL_BB_SQR_CODE \ |
---|
823 | lo=sqrtab[a&255];\ |
---|
824 | lo=lo|(sqrtab[(a>>8)&255]<<16);\ |
---|
825 | lo=lo|(sqrtab[(a>>16)&255]<<32);\ |
---|
826 | lo=lo|(sqrtab[(a>>24)&255]<<48);\ |
---|
827 | hi=sqrtab[(a>>32)&255];\ |
---|
828 | hi=hi|(sqrtab[(a>>40)&255]<<16);\ |
---|
829 | hi=hi|(sqrtab[(a>>48)&255]<<32);\ |
---|
830 | hi=hi|(sqrtab[(a>>56)&255]<<48);\ |
---|
831 | |
---|
832 | |
---|
833 | |
---|
834 | |
---|
835 | #define NTL_BB_REV_CODE (revtab[(a>>0)&255]<<56)\ |
---|
836 | |(revtab[(a>>8)&255]<<48)\ |
---|
837 | |(revtab[(a>>16)&255]<<40)\ |
---|
838 | |(revtab[(a>>24)&255]<<32)\ |
---|
839 | |(revtab[(a>>32)&255]<<24)\ |
---|
840 | |(revtab[(a>>40)&255]<<16)\ |
---|
841 | |(revtab[(a>>48)&255]<<8)\ |
---|
842 | |(revtab[(a>>56)&255]<<0) |
---|
843 | |
---|
844 | #define NTL_MIN_LONG (-NTL_MAX_LONG - 1L) |
---|
845 | #define NTL_MIN_INT (-NTL_MAX_INT - 1) |
---|
846 | #endif |
---|
847 | |
---|